Watch this week's video on YouTube

Imagine you need to join two tables of data and filter the results. Perhaps you also need to convert some of the values for display as well (eg. 0 => "No", 1 => "Yes").

Do you choose to perform all of this in your SQL query? Or do you bring the data into your app and and handle it there with code?

Option 1: Write it in a SQL query

Usually I'm inclined to do as much in the SQL query as possible because:

Optimization

Relational SQL databases tend to be optimized for doing things like joins, filtering, aggregations, etc….

Software engineers have spent decades tuning their sorting algorithms and squashing bugs to make their relational databases handle these operations efficiently. Sure, you can probably find the occasional edge case where you could handcraft some app code to accomplish one of these things faster, but in most cases it's not worth the additional time and effort.

Ordered Data

The types of operations in the example above (joins, filtering) benefit from ordered data.

Indexes in your database offer data stored in a predefined order, making all of those operations faster. If you need to perform a transformation that can utilize an index, it usually makes sense to let the database handle this operation in the SQL query rather than moving that data to your app and not have an index available.

Data Reduction

Joining and filtering your data in your database allows you to send a reduced number of records over the network to your application.

Network speeds tend to be one of the largest bottle necks in application systems, so eliminating the amount of data you need to pass through them up front in your SQL query can greatly improve the performance of your application.

This is especially true if you have switched to a work-from-home world where network speeds are even more of a bottleneck.

Hardware

This one can be a toss up. The SQL database servers I usually work with are beefy, having significantly more memory and CPU than my general purpose application servers (or my laptop running analytical code).

This means I can run most of my transformation logic faster on my database server than I can anywhere else. This comes at the cost of your performance hungry queries taking away resources from other queries that may be running on the database server at the same time, but on servers not running at capacity this trade off can be worth the speed. More on this in our app code option.

Portability

If you need to share your transformation logic, a SQL query is going to be easier to share with others or include in other processes than application code.

For example, SQL queries are like a universal language in most companies: programmers, analysts, data scientists, technical product managers, and anyone else who may be interested in understanding the business logic will be able to get a high-level understanding by looking at the SQL query. It will also be easier for them to incorporate into their processes.

Contrast that with some app code that is not easy to execute without installing dependencies and understanding language specific syntax; a SQL query will always be easier to share and reuse.

Option 2: Code it in the application

Everything you can write in a SQL query can be performed in most programming languages.

Let's discuss when it makes sense to store transformation logic in app code.

Domain Specific Performance Improvements

Most SQL relational databases are built to be general purpose. That's what makes them so powerful: they work well to help solve many different problems without needing domain specific optimizations.

Sometimes though, you may know something about your data that will allow you to work with it more efficiently than a relational database.

For example, maybe your data consists of mostly XML files, or you know you will be doing a lot of recursive processing of your data. Sure, most relational database engines have some way to accomplish these tasks, but performing this type of processing will most likely be faster to do in your application code.

Local Caching

Sometimes your application will need to process and reprocess the same data over and over again.

In these cases, it may make sense to transfer your data across the network a single time to your application, so it can locally process (and reprocess) that data as many times as needed.

Costs

Many enterprise database platforms cost money to license. Application servers usually don't have those same kinds of costs. That means that running your transformation logic in a SQL query on your database server is often more expensive than in code on an application server.

If your database server is not at full capacity, then this likely isn't an issue - after all you want to maximize the usage of the licenses you are paying for.

However, if your database environment is at or over capacity, offloading some transformation logic into your application code may make it run faster and will also likely be cheaper to run.

The cloud makes this case even more transparent, with database as a service offerings generally being more expensive than their application compute counterparts.

Conclusion: which is better?

It depends on your situation.

This post speaks to generalized scenarios. For specific uses, there will be times where it makes sense to store your logic in SQL queries, and other times in your application code.

The point is the next time before you dive in and start writing code, think through your restrictions and goals to figure out the best location to run things.

Watch this week's video on YouTube

A Giant Heap

Recently I had to filter out 1.2 billion records from a 3.5 billion row heap. Don't ask me why this 3.5 billion row table is stored as a heap.

If the lack of a clustered index wasn't bad enough, I also had some other restrictions:

• I couldn't add a clustered index (or any index for that matter) sorted on the key I needed to filter on. It wasn't my system, and I needed to access the 1.2 billion records sooner than it would take to get a clustered index approved and added. Even then, maybe the lack of a clustered index on this table is a feature and my request would be denied. Who knows.
• I didn't have a server that could store all of the data. At first I thought of copying all 3.5 billion rows to my own server and indexing it how I needed, but I didn't have enough storage space anywhere to do that.
• My connection to the server had a relatively short timeout set on it. This also couldn't be changed. If I couldn't copy all 3.5 billion rows because of storage, I also couldn't copy all 1.2 billion records in one fell swoop because the connection would timeout.

Iterative Process

I struggled with this problem for a little bit. My deadline clock was ticking and I was stuck as to how I could copy and subsequently query the 1.2 billion rows of data I needed. My focus transitioned from "what is the best way to do this" to "how do I do this".

The solution that ended up working for me was to query the table hundreds of times, each time filtering out and copying only 1 week of data by running a query similar to this:

SELECT *
FROM dbo.MyBigHeap
WHERE
    CreateDate >= @StartDate
    AND CreateDate < @StopDate

Yes, this did cause me to scan the entire table hundreds of times, but in the end it was the right amount of data that I could copy at a time before the connection timed out.

Eventually I had the 1.2 billion rows I needed copied to my own server. I had a clustered column store index on the table (primarily for the compression savings) and some nonclustered indexes to support the queries I would need to run on it. Was this the best solution? I don't know. But it worked for me given the constraints and deadline I had to meet.

Moral of the Story

Always put a clustered index on your tables. Even if you don't have a use case to sort/filter them immediately, you will be creating a world of pain when someone comes along who does need to query that data.

Watch this week's video on YouTube

Hey everyone. Long time no chat. I wanted to write this quick post to let you know why there haven't been any videos in a while. If you are interested in technical content, skip this post and come back next week for a post on filtering data from heaps.

First off, thanks to everyone who reached out to see if things were ok. Things are great! Well, as good as they can be anyway. Since my last post/video a few things have changed in my life:

• I got a new job building out a new data science team
• My wife and I had our second child, which has been very exciting
• COVID19 has turned the world upside down

Combine all of that with America's racial atrocities and injustices being put in the spotlight and well, writing new posts and filming videos just wasn't a priority.

And while the world still sometimes feels like a terrible place, and I'm still not getting adequate sleep with a 4 month old in the house, I've been wanting to get back to writing blog posts and making videos because they bring me, and I know many of you, enjoyment.

With that said, going forward I am going to make a few changes to make things more sustainable:

• For over 3 years, I wrote posts and filmed videos every week. Most of the time this was manageable, but sometimes it caused me stress to meet deadlines for something that I'm doing for fun. So while I still plan to update this blog regularly, I plan on doing it in a way that doesn't self-impose stress.
• You may have noticed the logo change to Data with Bert. SQL has always been just a small slice of my technical life and I've wanted to incorporate the rest of my data-related projects into my website. Things like what I'm currently building with Arduinos, and how I'm using the cloud to store data in non-relational services. Writing about these other projects never felt right given the site had "SQL" in the name, so that's what brought about the change. I am still going to write primarily about SQL, but I'm basically giving myself permission to widen the topic area to other data-related projects. They will all still have a heavy data problem solving focus, so if you've enjoyed the posts up to this point, I'm sure you will find the future posts relevant as well.

Anyway, thanks for reading :).

Watch this week's video on YouTube

This week, I want to share my process for analyzing Twitter. Specifically, I want to find who all of my friends follow on Twitter that I don't currently follow. Essentially, I want to build a Twitter follower recommendation engine.

Let's start with the theory behind the problem I'm trying to solve. On Twitter, I follow a bunch of people. Around 450 at the time of filming this video.

These are people I've either met in person and want to keep in touch with or am interested in following - or both!

Now I don't attend many conferences and events, so finding more people to follow that way is slow going; I probably only meet a handful of new people every year.

However, the second category of people, the interesting ones out there on the internet, is almost limitless. The problem is finding them in a virtual sea of bots, marketing-only accounts, and complainers.

Twitter does have a "Who to Follow" page, and while I'm sure it has some great suggestions, I don't necessarily trust all of Twitter's recommendations. It's kind of like trusting a site like Yelp for restaurant reviews when the local McDonalds rates better than your favorite burger place. I'm sure that particular McDonalds is beautiful by fast food standards, but I'm looking for more personalized recommendations.

So that leads me to this project: I decided to find better personalized recommendations on Twitter by looking at who my friends are following that I am not.

Getting the Data

To start, I needed a list of the people I follow on Twitter, and then a list of who they follow. The proper way to do this is to use the official Twitter API, so I wrote some Python code to do exactly that. Twitter calls the people you follow "Friends", which is funny since I usually don't consider one-way relationships friendships.

I won't go through the code in detail (you can download it from GitHub if you'd like to do that) but essentially it uses the friends/list endpoint to download all of my Friends. I then went through each of those Friends and found all of their Friends.

In total, this ended up being around 125,000 people.

As a quick side note, the Twitter API is terrible. Using it is fine, but the rate limiting is horrendous. I was basically capped at downloading 12,000 people per hour, which meant it took about a day to download all of the data I needed.

Graph Analysis with NetworkX

Once I had the data downloaded, it was time to find relationships between my friends and the people they follow. For this, I decided to use an open source Python library called NetworkX. NetworkX helps perform complex network analysis, which is perfect for what I was trying to do.

NetworkX uses a graph structure to help with its analysis. A graph is made up of of nodes and edges. In our case, the Twitter users are our nodes, and our edges are the relationships. The first thing I did was load all of the people I follow and created a directional edge (aka. an arrow) to indicate the one-way relationship from me to them. Next, I looped over all of those people's friends, adding additional nodes and edges.

At this point, my graph had 125,000 nodes which was way too many to draw quickly on my computer, and way too many to model with Lego minifigs. I figured not ALL of this data would be useful and I needed a way to filter it.

I started by filtering out any of my friends' friends who happen to already be my friends - this wouldn't be useful since I already follow these people.

Next, I decided to keep only the top 50 most followed people in that second tier of Twitter users. This would limit the processing needed while still giving me a list of the most followed Twitter users that my friends follow that I don't currently follow. Once again, all of this code is available on my GitHub.

At this point, I was able to use NetworkX's drawing capabilities to beautifully render the network of users and relationships. Sure, I could have just listed out the top users that I don't follow, but there is something special about being able to create a visual network of those relationships.

So that's it. It was fun using a graph library to find new people to follow on Twitter using the people I currently follow as a proxy for finding relevant users instead of Twitter's black box algorithms.

Watch this week's video on YouTube

Pop quiz: What will be the output of the below three statements?

DECLARE @val_varchar varchar(3) = '100';
DECLARE @val_bigint bigint = 100;
DECLARE @val_tinyint tinyint = 100;

PRINT('varchar to varchar:');
PRINT(CAST(@val_varchar AS VARCHAR(2)));

PRINT('bigint to varchar:');
PRINT(CAST(@val_bigint AS VARCHAR(2)));

PRINT('tinyint to varchar:');
PRINT(CAST(@val_tinyint AS VARCHAR(2)));

As you might have guessed, the first conversion truncates the value, leaving the answer as "10":

The second response is also somewhat intuitive - a number can't be converted into string that has fewer places than original digits:

However, the final answer returns:

NOTE: if we change all of the conversions to VARCHAR(3) above, all three results return the string value '100'

What the ****?

Why does SQL Server sometimes error when converting a number into a string, but other times succeeds and returns an asterisk?

I don't know.

The best (and logical) answer I could find online is from Robert Sheldon, who attributes it to poor error handling practices, "...before error handling got a more reputable foothold."

My understanding is that the asterisk was originally the default truncation value. Later, when the SQL Server development team was adding the bigint datatype, they decided throwing an exception is a better way to handle truncation in errors. However, probably for backwards compatibility, they never went back to change the behavior of the asterisks for pre-existing datatypes.

The documentation is clear about this behavior as well: any int, smallint, or tiny int when converted to char or varchar with fewer characters will result in an asterisk.

This becomes a problem if you create strings out of tinyints, smallints, and ints and suddenly start receiving values larger than you originally expected:

DECLARE @Age int = 75;
PRINT(CAST(@Age AS VARCHAR(2)) + ' years old.');

SET @Age = 100;
PRINT(CAST(@Age AS VARCHAR(2)) + ' years old.');

Conversions are hard

So is SQL Server to blame for this issue? Sure. But I can empathize: breaking changes are not something people typically want to add to a product.

I've written before how implicit conversions can lead to funny results in SQL Server. This asterisk scenario is another case of SQL Server having to make a judgement call about handling an impossible situation you created for it.

While it's easy to complain about whether SQL Server handles the conversion the correct way or not, at the end of the day the responsibility falls on you to prevent these types of impossible calculations from being requested of SQL Server in the first place.